fix: possible unsent function call in the last chunk of streaming response in OpenAI provider

[bowenliang123]

• fix possible unsent function call in the last chunk of streaming response in OpenAI provider
• the delta_assistant_message_function_call_storage caches the function calls when function_call is found and then it continues to process the next chuck. However, if the last message itself contains a full function calling, it's skipped and never sent
• the corner case comes with using OpenAI provider to call the LLMs under Open-compatible proxies, like we have GLM3 and Qwen

[bowenliang123]

cc @Yeuoly @takatost

[Yeuoly]

BTW, openai_api_compatible will support function call soon, you can then use it as your proxies of GLM3 and Qwen.

Thanks! best wishes to your new year.

[bowenliang123]

Thanks. We probably continue to use the openai provider to connect the LLMs via the openai compatible proxies.

1. for the earlier support of new features, Dify often consider OpenAI services first.
2. a set of predefined models of OpenAI helps us to provide all the selected services to out internal users, instead of forcing them to add the LLMs one by one in every workspace.

I raised this PR against OpenAI provider, for the reason that it does wrongly skipped the last message of function call. Especially in the future, OpenAI models may be able to return the arguments with the function calling at one message.

And Happy new Year!